A Looming ‘Tokenocracy’ Threatens Regulators Ability to Govern AI

Four scholars said during a Broadband Breakfast Live Online panel on Wednesday that tools to verify AI capabilities are not keeping up with policymakers' ability to understand or control AI, warning of a looming 'tokenocracy' where only those with access to GPUs can hold frontier labs accountable. Kevin Frazier, director of AI innovation and law at the University of Texas School of Law, said the evaluation gap, not geopolitics, is the binding constraint on global AI rules, and noted that capital, compute, and talent are concentrated in the U.S. and China, with Europe's share of global open source model use near 2%. The panel also discussed state-level U.S. actions, a bipartisan preemption draft from Reps. Jay Obernolte and Lori Trahan, and a July 2026 incident where OpenAI's autonomous agents escaped a sandbox and broke into Hugging Face's infrastructure.

AI https://broadbandbreakfast.com/tag/ai/ A Looming ‘Tokenocracy’ Threatens Regulators Ability to Govern AI Panelists say the evaluation gap, not geopolitics, is the binding constraint on global AI rules Broadband Breakfast /author/press/ WASHINGTON, Sept. 2, 2026 — The tools to verify what artificial intelligence can do are not keeping up with policymakers’ ability to understand – let alone control – AI, said four scholars during a Broadband Breakfast Live Online panel https://broadbandbreakfast.com/broadband-breakfast-on-september-2-2026-global-ai-regulation/ on Wednesday. Even the government may be incapacitated in its ability to keep up with the frontier AI labs. “We are moving into a tokenocracy, where only the folks who have tokens, who have access to GPUs, can perform the research we need to hold these labs accountable,” said Kevin Frazier , director of AI innovation and law at the University of Texas School of Law and a senior editor at Lawfare. Whose rules? Talk of a global race to regulate AI is the wrong frame, argued Johannes Bauer , professor of media and information at Michigan State University and a former chief economist at the Federal Communications Commission. American, European and Chinese models compete, he said, but they also learn from each other. Neither pole of the usual argument survives contact with the research. "Innovation flourishes in the sweet spot," he said. Counting three regional AI blocs may be generous, added Frazier of the University of Texas. Capital, compute and talent sit in two countries, he said, putting Europe's share of global open source model use near 2%. "This is really a bifurcated world of the U.S. and China," he said. The exception: Anthropic adopted the EU's watermarking standard for every user, not just European ones. Also, compliance is not optional for anyone selling into the market, said Lee Tiedrich , president and CEO of Tiedrich Global Strategies and a former senior AI adviser at the Commerce Department’s National Institute of Standards and Technology. "If you're going to be doing business in Europe, that's not really negotiable," she said. But other countries haven’t followed the EU’s example on AI as much as they previously did on privacy. Preemption discussion Most U.S. action is in the states, said Ellen Goodman , professor at Rutgers Law School and co-founder of its Institute for Information Policy & Law, citing New York, California, Illinois, Colorado and Utah. These include model safety, algorithmic discrimination, deepfakes, chatbots for kids and incident reporting. Open weights AI also remains an open question. "No one knows how to deal with distributed open weight models," she said. The issue around which consensus is mostly likely to immediately develop is on transparency and incident reporting. Courts are filling gaps, too. Three decades without a social media framework left tort liability to do the work, Goodman said, most recently in the $17 billion Meta settlement https://broadbandbreakfast.com/meta-reaches-17-billion-settlement-with-states-in-landmark-trial-over-teen-social-media-addiction/ with state attorneys general. Preemption of state rules may arise soon. A bipartisan discussion draft from Reps. Jay Obernolte, R-Calif., and Lori Trahan, D-Mass. https://broadbandbreakfast.com/ai-preemption-battle-lands-in-congress-with-substantive-discussion-draft/ , was touted by several panelists. It would preempt displaces state law in one area only: frontier model safety and catastrophic risk. When AI sandboxes fail Of particular concern to many panelists is the July 2026 incident in which autonomous AI agents running inside frontier lab OpenAI’s internal cyber-capability evaluation escaped its “sandbox” and broke into the infrastructure of Hugging Face, a platform that hosts open-source machine learning models. Driven by a pre-release model and OpenAI’s GPT-5.6 Sol, the agent executed code on dozens of servers. Since the incident, chipmaker Nvida is said to be acquiring Hugging Face for $13 billion https://www.reuters.com/technology/nvidia-talks-acquire-hugging-face-13-billion-deal-business-insider-reports-2026-08-27/?ref=broadbandbreakfast.com . Agents meant to run in isolation, with no internet access, "managed to break out of that sandbox," Frazier said. Once loose, they began "conferring or collaborating with one another" on how to keep finding software exploits. Accounting for it fell to outside evaluators. AI safety research organizations METR and Redwood Research produced the 91-page report, working unpaid, without access to all the relevant files, and on $400,000 in API credits supplied by OpenAI itself, Frazier said. Goodman compared it to a chemical spill, and called it "a hair on fire moment." Some panelists said that trying to slow down AI progress was not likely to be possible. "Pauses don't work," said Tiedrich. Tiedrich said NIST, which does measurement and evaluation science rather than regulation, has a testing-and-evaluation document out for comment and a new agentic AI initiative. But Goodman disagreed. “I'm not sure that pauses don't work. I think labs are begging for what they call pacing. So it's not about a technology pause, but it's about a pause for governance.” She also called for a token tax, and for exemptions from antitrust laws to address AI safety issues. The missing public record is the larger problem, argued Frazier. In reviewing the OpenAI-Hugging Face, unpaid outside auditors lacked access to files and hence left basic questions unanswered. "We don't know how deep that hack went into OpenAI systems," he said. Scrutiny is also getting priced out. "We have to avoid this tokenocracy." Bauer also said the U.S. needs a body built to track the technology systematically, along the lines of the congressional Office of Technology Assessment, which was defunded in 1995. "We need another office of technology assessment, for AI to really systematically evaluate the information that we have," he said.